Feature/trim compatible

Project.toml

@@ -1,7 +1,7 @@
 name = "HostCPUFeatures"
 uuid = "3e5b6fbb-0976-4d2c-9146-d79de83f2fb0"
 authors = ["Chris Elrod <elrodc@gmail.com> and contributors"]
-version = "0.1.17"
+version = "0.1.18"
 
 [deps]
 BitTwiddlingConvenienceFunctions = "62783981-4cbd-42fc-bca8-16325de8dc4b"

src/HostCPUFeatures.jl

@@ -1,7 +1,6 @@
 module HostCPUFeatures
-if isdefined(Base, :Experimental) &&
-   isdefined(Base.Experimental, Symbol("@max_methods"))
-    @eval Base.Experimental.@max_methods 1
+if isdefined(Base, :Experimental) && isdefined(Base.Experimental, Symbol("@max_methods"))
+  @eval Base.Experimental.@max_methods 1
 end
 
 using Libdl, Static
@@ -10,8 +9,13 @@ using IfElse: ifelse
 
 using BitTwiddlingConvenienceFunctions: prevpow2, nextpow2, intlog2
 
-export has_feature, fma_fast, pick_vector_width, pick_vector_width_shift, register_count,
-  register_size, simd_integer_register_size
+export has_feature,
+  fma_fast,
+  pick_vector_width,
+  pick_vector_width_shift,
+  register_count,
+  register_size,
+  simd_integer_register_size
 
 function get_cpu_name()::String
   if isdefined(Sys, :CPU_NAME)
@@ -22,19 +26,19 @@ function get_cpu_name()::String
 end
 include("cpu_info.jl")
 if (Sys.ARCH === :x86_64) || (Sys.ARCH === :i686)
-    include("cpu_info_x86.jl")
+  include("cpu_info_x86.jl")
 elseif Sys.ARCH === :aarch64
-    include("cpu_info_aarch64.jl")
+  include("cpu_info_aarch64.jl")
 else
-    include("cpu_info_generic.jl")
+  include("cpu_info_generic.jl")
 end
 include("pick_vector_width.jl")
+include("static_features.jl")
 
 unwrap(::Val{S}) where {S} = S
 unwrap(::StaticInt{S}) where {S} = S
 unwrap(::StaticFloat64{S}) where {S} = S
 unwrap(::StaticSymbol{S}) where {S} = S
-
 @noinline function redefine()
   @debug "Defining CPU name."
   define_cpu_name()
@@ -43,14 +47,14 @@ unwrap(::StaticSymbol{S}) where {S} = S
   reset_extra_features!()
 end
 const BASELINE_CPU_NAME = get_cpu_name()
-function __init__()
-  ccall(:jl_generating_output, Cint, ()) == 1 && return
-  if Sys.ARCH === :x86_64 || Sys.ARCH === :i686
-    target = Base.unsafe_string(Base.JLOptions().cpu_target)
-    occursin("native",  target) || return make_generic(target)
-  end
-  BASELINE_CPU_NAME == Sys.CPU_NAME::String || redefine()
-  return nothing
-end
+# function __init__()
+#     ccall(:jl_generating_output, Cint, ()) == 1 && return
+#     if Sys.ARCH === :x86_64 || Sys.ARCH === :i686
+#         target = Base.unsafe_string(Base.JLOptions().cpu_target)
+#         occursin("native", target) || return make_generic(target)
+#     end
+#     BASELINE_CPU_NAME == Sys.CPU_NAME::String || redefine()
+#     return nothing
+# end
 
 end

src/cpu_info.jl

@@ -1,10 +1,15 @@
 function feature_string()
-    llvmlib_path = VERSION ≥ v"1.6.0-DEV.1429" ? Base.libllvm_path() : only(filter(lib->occursin(r"LLVM\b", basename(lib)), Libdl.dllist()))
-    libllvm = Libdl.dlopen(llvmlib_path)
-    gethostcpufeatures = Libdl.dlsym(libllvm, :LLVMGetHostCPUFeatures)
-    features_cstring = ccall(gethostcpufeatures, Cstring, ())
-    features = filter(ext -> (ext ≠ "" && (m = match(r"\d", ext); isnothing(m) ? true : m.offset != 2)), split(unsafe_string(features_cstring), ','))
-    features, features_cstring
+  llvmlib_path =
+    VERSION ≥ v"1.6.0-DEV.1429" ? Base.libllvm_path() :
+    only(filter(lib -> occursin(r"LLVM\b", basename(lib)), Libdl.dllist()))
+  libllvm = Libdl.dlopen(llvmlib_path)
+  gethostcpufeatures = Libdl.dlsym(libllvm, :LLVMGetHostCPUFeatures)
+  features_cstring = ccall(gethostcpufeatures, Cstring, ())
+  features = filter(
+    ext -> (ext ≠ "" && (m = match(r"\d", ext); isnothing(m) ? true : m.offset != 2)),
+    split(unsafe_string(features_cstring), ','),
+  )
+  features, features_cstring
 end
 
 const FEATURE_SET = Set{String}()
@@ -16,20 +21,20 @@ archstr() = Sys.ARCH === :i686 ? "x86_64_" : string(Sys.ARCH) * '_'
 feature_name(ext) = archstr() * ext[2:end]
 process_feature(ext) = (feature_name(ext), first(ext) == '+')
 
-has_feature(_) = False()
-@noinline function set_feature(feature::String, has::Bool)
-    featqn = QuoteNode(Symbol(feature))
-    if has
-        @eval has_feature(::Val{$featqn}) = True()
-    else
-        @eval has_feature(::Val{$featqn}) = False()
-    end
-end
+# has_feature(_) = False()
+# @noinline function set_feature(feature::String, has::Bool)
+#     featqn = QuoteNode(Symbol(feature))
+#     if has
+#         @eval has_feature(::Val{$featqn}) = True()
+#     else
+#         @eval has_feature(::Val{$featqn}) = False()
+#     end
+# end
 
 function set_features!()
   features, features_cstring = feature_string()
   znver3 = get_cpu_name() === "znver3"
-  for ext ∈ features
+  for ext in features
     feature, has = process_feature(ext)
     if znver3 && occursin("512", feature)
       has = false
@@ -39,27 +44,26 @@ function set_features!()
   end
   Libc.free(features_cstring)
 end
-set_features!()
-
+# set_features!()
 
 
 function reset_features!()
-    features, features_cstring = feature_string()
-    for ext ∈ features
-        feature, has = process_feature(ext)
-        if _has_feature(feature) ≠ has
-            @debug "Defining $(has ? "presence" : "absense") of feature $feature."
-            set_feature(feature, has)
-        end
+  features, features_cstring = feature_string()
+  for ext in features
+    feature, has = process_feature(ext)
+    if _has_feature(feature) ≠ has
+      @debug "Defining $(has ? "presence" : "absense") of feature $feature."
+      set_feature(feature, has)
     end
-    Libc.free(features_cstring)
+  end
+  Libc.free(features_cstring)
 end
 
 register_size(::Type{T}) where {T} = register_size()
 register_size(::Type{T}) where {T<:Union{Signed,Unsigned}} = simd_integer_register_size()
 
 function define_cpu_name()
-    cpu = QuoteNode(Symbol(get_cpu_name()))
-    @eval cpu_name() = Val{$cpu}()
+  cpu = QuoteNode(Symbol(get_cpu_name()))
+  @eval cpu_name() = Val{$cpu}()
 end
 define_cpu_name()

src/cpu_info_aarch64.jl

@@ -1,10 +1,12 @@
 
-_has_aarch64_sve() = (Base.libllvm_version ≥ v"11") && (Base.BinaryPlatforms.CPUID.test_cpu_feature(Base.BinaryPlatforms.CPUID.JL_AArch64_sve))
+_has_aarch64_sve() =
+  (Base.libllvm_version ≥ v"11") &&
+  (Base.BinaryPlatforms.CPUID.test_cpu_feature(Base.BinaryPlatforms.CPUID.JL_AArch64_sve))
 
 if Int === Int64
-    @noinline vscale() = ccall("llvm.vscale.i64", llvmcall, Int64, ())
+  @noinline vscale() = ccall("llvm.vscale.i64", llvmcall, Int64, ())
 else
-    @noinline vscale() = ccall("llvm.vscale.i32", llvmcall, Int32, ())
+  @noinline vscale() = ccall("llvm.vscale.i32", llvmcall, Int32, ())
 end
 
 # TODO: find actually support SVE
@@ -20,30 +22,30 @@ function _dynamic_register_size()
 end
 
 function _set_sve_vector_width!(bytes = _dynamic_register_size())
-    @eval begin
-        register_size() = StaticInt{$bytes}()
-        simd_integer_register_size() = StaticInt{$bytes}()
-    end
-    nothing
+  @eval begin
+    register_size() = StaticInt{$bytes}()
+    simd_integer_register_size() = StaticInt{$bytes}()
+  end
+  nothing
 end
 
 
 if _has_aarch64_sve()# && !(Bool(has_feature(Val(:aarch64_sve))))
-    has_feature(::Val{:aarch64_sve_cpuid}) = True()
-    _set_sve_vector_width!()
+  has_feature(::Val{:aarch64_sve_cpuid}) = True()
+  _set_sve_vector_width!()
 else
-    # has_feature(::Val{:aarch64_svejl}) = False()
-    register_size() = StaticInt{16}()
-    simd_integer_register_size() = StaticInt{16}()
+  # has_feature(::Val{:aarch64_svejl}) = False()
+  register_size() = StaticInt{16}()
+  simd_integer_register_size() = StaticInt{16}()
 end
 
 function reset_extra_features!()
-    drs = _dynamic_register_size()
-    register_size() ≠ drs && _set_sve_vector_width!(drs)
-    hassve = _has_aarch64_sve()
-    if hassve ≠ has_feature(Val(:aarch64_sve_cpuid))
-        @eval has_feature(::Val{:aarch64_sve_cpuid}) = $(Expr(:call, hassve ? :True : :False))
-    end
+  drs = _dynamic_register_size()
+  register_size() ≠ drs && _set_sve_vector_width!(drs)
+  hassve = _has_aarch64_sve()
+  if hassve ≠ has_feature(Val(:aarch64_sve_cpuid))
+    @eval has_feature(::Val{:aarch64_sve_cpuid}) = $(Expr(:call, hassve ? :True : :False))
+  end
 end
 
 fma_fast() = True()
@@ -53,4 +55,3 @@ has_opmask_registers() = has_feature(Val(:aarch64_sve_cpuid))
 fast_int64_to_double() = True()
 
 fast_half() = False()
-

src/cpu_info_generic.jl

@@ -9,4 +9,3 @@ reset_extra_features!() = nothing
 
 fast_int64_to_double() = True()
 fast_half() = False()
-